Samsung's contract chip business is reportedly closing in on two of the biggest names in artificial intelligence. According to reports circulating this week, Samsung Foundry is in talks to manufacture custom 2nm AI chips for both Anthropic and Meta — deals that, if signed, would mark a meaningful win over rival TSMC as demand for purpose-built AI silicon explodes. Both companies are chasing custom accelerators to cut their reliance on expensive, general-purpose Nvidia GPUs, and Samsung's cutting-edge 2nm process is at the center of the conversations.
Key Takeaways
- Samsung Foundry is reportedly in talks to build 2nm AI accelerator chips for both Anthropic and Meta.
- Meta's third-generation MTIA chips could reportedly be worth more than ₩10 trillion (around ₱400 billion), shifting work away from TSMC.
- Anthropic is planning roughly 1GW of data-center capacity backed by about $50 billion (around ₱3.1 trillion), much of it hardware.
- Samsung's 2nm SF2 node entered volume production in late 2025 using Gate-All-Around transistors.
- Samsung already backed Anthropic's $65 billion (around ₱4 trillion) Series H round in May 2026.
Samsung Foundry reportedly courting Anthropic and Meta
The reports come from two threads. The Information first reported that Anthropic was exploring Samsung's manufacturing for a custom chip, while South Korea's SEDaily reported that Samsung Foundry could produce Meta's next-generation AI accelerator. The common denominator is Samsung's 2nm process technology, the same leading-edge node Samsung is pushing to win high-margin AI work. Both sets of talks are described as early and unconfirmed by the companies involved, so nothing is finalized.
Meta's MTIA chips and the move away from TSMC
For Meta, the prize is its third-generation MTIA (Meta Training and Inference Accelerator). Samsung could reportedly manufacture those chips in a deal valued at more than ₩10 trillion (roughly $6.5 billion, or around ₱400 billion). That would represent a notable shift away from TSMC, which built Meta's earlier MTIA generations but is stretched thin by demand from major clients including Apple and Nvidia. Meta is racing to stand up AI data centers with a combined 5GW of capacity by 2030 and has reportedly targeted a new chip generation roughly every six months, a cadence that requires committed foundry capacity.
Anthropic's custom-chip ambitions
Anthropic's plans are just as aggressive. The company is reportedly evaluating Samsung's 2nm process and advanced packaging for a custom inference ASIC, and is planning roughly one gigawatt of total data-center capacity backed by approximately $50 billion (around ₱3.1 trillion) in investment — with about half earmarked for physical hardware such as processors, DRAM, and storage. The strategic intent looks serious: Anthropic recently hired Clive Chan, who left OpenAI in June after helping build that company's custom-chip program.
There is already a financial thread tying the two together. Samsung became a strategic infrastructure partner during Anthropic's $65 billion (around ₱4 trillion) Series H round in May 2026, a raise that also drew in memory makers SK Hynix and Micron. That relationship is part of what makes a manufacturing deal plausible, and it fits a wider pattern of AI labs and chip suppliers entangling their balance sheets, echoing moves like Apple's lobbying to source memory chips from China's blacklisted CXMT.
Why 2nm matters
Samsung's 2nm family is its first to use Gate-All-Around (GAA) transistors, an architecture that wraps the gate around the channel on all sides for better efficiency than older FinFET designs. The SF2 node reportedly entered volume production in late 2025, while a more advanced SF2Z variant with backside power delivery targets mass production in 2027. Custom inference chips built on such nodes can strip out the general-purpose overhead that GPUs carry: OpenAI's competing Jalapeño inference chip has reportedly shown roughly 50% lower inference costs than GPU-based setups in early testing, which is exactly the kind of saving that justifies designing bespoke silicon.
The bigger picture
Even unconfirmed, the reports underline how the AI arms race is moving down the stack into hardware — and how much a foundry win matters. The Anthropic thread emerged earlier in the week, with the Meta angle expanding the story on July 4-5, and both remain reports rather than announcements. For Samsung, landing even one of these customers would validate its 2nm push against TSMC. And it is not the only AI player probing Samsung's silicon ambitions; the company's chip decisions increasingly ripple across the whole industry, as seen in coverage of Anthropic's own access and infrastructure fights.
Sources:
